Key takeaway
- Edge analytics runs analysis next to sensors, machines and cameras. Learn when it fits, how the pipeline works, current 2027 platforms and the trade-offs.
Edge analytics is the practice of analysing data on or near the device that produces it, such as a gateway, an industrial PC or a camera, so that a decision can be made locally in seconds rather than after a round trip to the cloud. The device or gateway filters, aggregates and scores the data, acts on what matters, and sends only summaries, alerts and exceptions to central platforms.
This guide focuses on the analytics step: when to run analysis at the edge, how the pipeline works, which platforms are current in 2026 and where the approach does not pay off. For how edge data is collected, stored and synchronised, see our guide to edge data management. For the wider architecture, see intelligent edge types and use cases and the difference between cloud, edge and fog computing.
What is edge analytics?
Edge analytics runs analytical logic, from simple threshold rules to trained machine learning models, close to sensors, machines and network devices. Microsoft describes the aim of its edge runtime in these terms: not all analytics need to be in the cloud, and running anomaly detection at the edge lets a team respond as quickly as possible to problems on a production line, while cleaning and aggregating data locally avoids transferring terabytes of raw data (Microsoft Learn, Azure IoT Edge).
It differs from edge data management in purpose. Edge data management is about moving and governing data reliably. Edge analytics is about turning that data into a decision on site: stop a machine, raise a quality alert, flag an intruder, or forward a summary for fleet-wide analysis. Our overview of IoT edge analytics trends covers sector examples in more depth.
How is edge analytics different from cloud analytics?
| Aspect | Edge analytics | Cloud analytics |
| Where it runs | Device, gateway or on-site cluster | Central data platform or warehouse |
| Typical question | What should happen at this asset now? | What patterns hold across all sites over time? |
| Data used | Live streams from local devices | Historical and cross-site data |
| Connectivity need | Can keep working when the link is down | Needs data to arrive centrally first |
| Compute | Constrained; models must be small and efficient | Elastic; suited to training and large joins |
| Best for | Control loops, alerts, filtering, inference | Model training, reporting, planning |
Most enterprises need both. A common split is to score and act at the edge, then send results and samples to the cloud to retrain models and compare sites. Our comparison of edge computing and cloud computing covers the infrastructure side of that choice.
When should you run analytics at the edge?
Microsoft's guidance for Stream Analytics on IoT Edge names four scenarios: low-latency command and control, limited connectivity, limited bandwidth, and regulatory compliance where sensitive data is processed on site and only compliant results go to the cloud (Microsoft Learn, Stream Analytics on IoT Edge). Those four tests make a practical decision table.
| Driver | When edge analytics fits | When the cloud is enough |
| Latency | A machine, line or safety system must react before a cloud round trip completes | Decisions can wait minutes or hours |
| Connectivity | Sites such as mines, rigs, ships or remote plants lose their link | Sites have reliable, redundant networks |
| Bandwidth | High-volume streams such as video or vibration data cost too much to send raw | Data volumes are small |
| Compliance | Raw data, such as video of people, should stay on site | No residency or privacy constraint on raw data |
How does an edge analytics pipeline work?
The steps are similar across vendors. Each step is a design decision about what runs locally and what moves to the centre.
- Connect devices. Sensors, PLCs and cameras publish to a local broker or gateway, typically over MQTT or OPC UA. Azure IoT Operations, for example, includes an edge-native MQTT broker and a connector for OPC UA (Microsoft Learn).
- Filter and aggregate. Drop noise, downsample and compute windowed statistics. AWS IoT Greengrass stream manager lets components read and write streams to filter, aggregate and analyse data on the core device before export (AWS documentation).
- Score. Run rules or a trained model on the stream. Runtimes such as LiteRT (formerly TensorFlow Lite), which accepts models from PyTorch, JAX and TensorFlow (Google), and ONNX Runtime, which targets CPU, GPU and NPU across cloud, edge, web and mobile (ONNX Runtime), are common choices.
- Act locally. Trigger an alarm, stop a machine, open a work order or flag a frame for review.
- Buffer and export. Send summaries, alerts and samples to the cloud, with priorities and bandwidth limits for when the link is weak. Greengrass stream manager exports to Amazon Kinesis Data Streams, AWS IoT SiteWise and Amazon S3; Azure IoT Operations data flows can send to Event Hubs or Kafka, Azure Data Lake Storage, Microsoft Fabric OneLake and Azure Data Explorer.
- Learn centrally. Retrain models on pooled data and push new versions back to the fleet.
Which edge analytics platforms are current going into 2027?
Several products named in earlier edge analytics guides, including this one, have since been retired, renamed or replaced. Check vendor status before you shortlist anything.
| Platform or runtime | What it does for edge analytics | Status (checked 11 November 2026) |
| AWS IoT Greengrass V2 | Edge runtime that lets devices act on local data, run ML predictions, and filter and aggregate data | Current. Greengrass V1 support ended on 7 October 2026 and its APIs are no longer available (AWS) |
| AWS IoT Analytics | Former cloud analytics destination for IoT data | Discontinued on 15 December 2025; Greengrass exports to it no longer function (AWS) |
| Azure IoT Operations | Unified data plane for the edge on Azure Arc-enabled Kubernetes, with MQTT broker, data flows and anomaly detection examples | Current. Can operate offline for a maximum of 72 hours (Microsoft) |
| Azure IoT Edge | Device runtime for containerised modules, including Stream Analytics and Azure Machine Learning | 1.6 LTS is the supported release; 1.5 LTS support ended on 10 November 2026 (Microsoft) |
| Azure Stream Analytics on IoT Edge | Streaming queries for command and control, filtering and aggregation near devices | Documented as current; Microsoft notes the older base images have reached end of life and points to updated images |
| LiteRT | On-device inference runtime | Current; formerly TensorFlow Lite |
| ONNX Runtime | Cross-platform inference engine with hardware acceleration | Current |
| Google Cloud IoT Core | Former managed device connection service | Retired on 16 August 2023 (Google product page, archived) |
What are the main use cases for edge analytics?
Predictive maintenance and anomaly detection
Microsoft's reference example connects an OPC UA asset to the edge MQTT broker, uses a data flow to normalise data before identifying anomalies, and sends processed data to Microsoft Fabric for shop-floor dashboards. Microsoft notes that predictive analytics with edge data helps anticipate failures before they occur and reduces downtime and maintenance costs.
Operational equipment effectiveness
Capturing and processing real-time data at the edge allows plants to monitor availability, performance and quality, and to find patterns that show where to improve. Where these signals need to feed decisions across robots, machines and on-site AI, NexaStack is an agentic operating platform for physical AI that runs across on-premises, cloud and hybrid environments.
Video analytics and smart surveillance
Video is the clearest case for analysing at the edge, because raw streams are large and often contain personal data. Scoring frames on site and forwarding only events keeps bandwidth and privacy exposure down. Our post on agentic AI for real-time video analytics explains how agents can act on those events.
Video: how real-time video analytics turns camera feeds into a live operations picture.
Retail and remote sites
Stores, warehouses and remote assets can combine camera, sensor and point-of-sale signals locally, then forward aggregates for central planning. Where the main constraint is a weak or costly link, edge analytics keeps the site working when the connection drops.
What has changed in edge analytics since this guide was first written?
- Platform churn. Google Cloud IoT Core and AWS IoT Analytics are gone, Greengrass V1 has ended and Azure IoT Edge 1.5 is out of support. Edge strategies built on those services need migration plans.
- Kubernetes at the edge. Azure IoT Operations runs as Kubernetes-native services on Arc-enabled clusters, which brings edge analytics closer to how platform teams already run cloud workloads.
- Model portability. LiteRT accepting PyTorch, JAX and TensorFlow models, and ONNX Runtime targeting CPU, GPU and NPU, reduce lock-in between training and edge inference.
- Agents at the edge. Teams now ask edge systems to take bounded actions, not only raise alerts. See agentic AI in edge computing.
How should you govern, secure and cost edge analytics?
Governance. Treat every edge model as a versioned asset with an owner, a known training dataset and a rollback path. Record which version made each decision so results can be explained later. Asset and schema registries, such as the Device Registry and schema registry in Azure IoT Operations, help keep data definitions consistent across sites. Once edge decisions start to drive agent actions, ElixirData Context OS keeps governed context, versioned policy and decision traces in one decision layer above AI agents.
Security. Edge devices sit in physically exposed places. Use device identity, managed secrets and certificates (Azure IoT Operations includes secrets management, certificate management and secure settings), signed updates and least-privilege network paths back to the cloud.
Cost. The savings come from sending less data and avoiding downtime. The costs come from hardware, fleet operations and model updates at many sites. Model both before scaling beyond a pilot. On the central side, AgentAnalyst lets teams ask questions of enterprise data in plain language and get real-time insights, which helps operations leaders use the summaries that edge sites send back.
What should you plan for in 2027?
Edge fleets depend on the operating system under the analytics runtime, and Microsoft's support page lists two 2027 dates.
- June 2027: end of OS provider standard support for Ubuntu Server 22.04, listed for both Tier 1 (AMD64, ARM64) and Tier 2 (ARM32v7) platforms (Microsoft Learn). Plan OS upgrades across sites before then.
- April 2027: the same page lists Ubuntu Core 1 with an April 2027 OS support date for Tier 1 platforms. Inventory any gateways that run it.
The page also lists Azure IoT Edge 1.6 LTS support to 14 November 2028, so the runtime has room while the operating system may not.
What adoption sequence works for edge analytics?
- Pick one decision. Choose a single decision where latency, bandwidth, connectivity or compliance clearly favours the edge.
- Prove it with rules first. Start with thresholds and windowed statistics before adding ML.
- Add a small model. Train centrally, convert for the edge runtime and measure accuracy on site.
- Design the return path. Decide which summaries, alerts and samples go back, with what priority.
- Industrialise. Add fleet deployment, monitoring, model versioning and rollback before scaling to more sites.
What are the trade-offs, and when should you not use edge analytics?
- Limited compute. Edge hardware constrains model size; complex models may need cloud scoring.
- Operational load. Hundreds of sites mean hundreds of runtimes to patch, monitor and secure.
- Fragmented views. Local decisions can drift between sites unless models and rules are managed centrally.
- Offline limits. Platforms have bounds; Azure IoT Operations, for example, documents a maximum of 72 hours offline, with possible degradation.
Do not use edge analytics when decisions can wait, when sites have reliable bandwidth and no data residency constraint, or when the question needs data from many sites at once. In those cases, central analytics is simpler and cheaper to run.
To map one edge decision your agents should make, and the context and policy it depends on, book a Context Workshop. See what happens in the workshop.
Frequently asked questions about edge analytics
What is edge analytics in simple terms?
It is analysis that runs on or near the device that produces the data, such as a gateway or camera, so decisions can be made locally and only summaries or alerts go to the cloud.
When does edge analytics make more sense than cloud analytics?
When a decision needs low latency, the site has limited or intermittent connectivity, bandwidth is costly, or regulation requires sensitive raw data to stay on site.
Which edge analytics platforms are current going into 2027?
AWS IoT Greengrass V2, Azure IoT Operations and Azure IoT Edge 1.6 LTS are current, with runtimes such as LiteRT and ONNX Runtime for on-device inference. Greengrass V1, AWS IoT Analytics and Google Cloud IoT Core have been retired.
What happened to TensorFlow Lite?
Google now calls it LiteRT. It is Google's inference runtime for edge devices and accepts models from PyTorch, JAX and TensorFlow.
Can edge analytics work without an internet connection?
Yes, within limits. Greengrass stream manager is designed for intermittent or limited connectivity, and Azure IoT Operations can operate offline for up to 72 hours.
Next Steps with Edge Computing
Talk to our experts about implementing compound AI system, How Industries and different departments use Agentic Workflows and Decision Intelligence to Become Decision Centric. Utilizes AI to automate and optimize IT support and operations, improving efficiency and responsiveness.
Follow XenonStack on LinkedIn: XenonStack on LinkedIn for new guides on agentic AI, data and platform engineering.
Ready to put AI agents to work? Explore ElixirClaw, the agentic OS platform to build, orchestrate and govern AI agents.
Explore the XenonStack platform family: ElixirData · ElixirClaw · ElixirHub · NexaStack · MetaSecure · Neural AI · AgentGRC · AgentSRE · AgentAnalyst · AgentInstruct · AgentRAI · AgentScrum · AgentSearch · AgentForce · AI Aviator · SAIF Aviator · CohortIQ · XenonStack AI · XenonStack UAE